Security Policy

shieldkit provides basic guardrails for AI SDK applications. It is not a complete security product. Read limitations before relying on it for compliance or threat prevention.

Supported versions

Version	Supported
0.1.x	Yes

Reporting a vulnerability

If you believe you have found a security issue in shieldkit:

1. Do not open a public GitHub issue for exploitable vulnerabilities.
2. Contact the repository maintainer privately with:
   • A description of the issue
   • Steps to reproduce
   • Impact assessment
   • Affected versions
3. Allow reasonable time for a fix before public disclosure.

Use GitHub Security → Advisories → Report a vulnerability when available.

Scope

In scope

• Bypass of input/output guardrails when used as documented
• Session budget or audit logging flaws
• Tool guard bypass (guardTools)
• Structured output repair leaking sensitive data unexpectedly

Out of scope

• Prompt injection that defeats regex/heuristic guards (documented limitation)
• Provider API security
• Consumer misconfiguration (disabled guards, missing sessionId, etc.)

Production recommendations

1. Use strict mode with includePartialInRetry: false when outputs may contain PII
2. Combine with provider-level safety features
3. Route audit logs to a retained, access-controlled sink
4. Treat serverless budgets as best-effort — see limitations
5. Require human approval for high-risk tools

Related

• Limitations
• Trade-offs
• Verification matrix